Predictive Uncertainty in State-Estimation Drives Active Sensing in Weakly Electric Fish

Abstract

AbstractAnimals adapt their sensory systems to accurately perceive their environment and therefore achieve optimal behavioral performance under varying sensory conditions. In addition to sensory adaptation, we propose that animals employ active sensing movements to shape the spatial and temporal characteristics of sensory signals. However, the mechanisms governing the generation of active sensing movements within the neural circuits of the central nervous system are not known. To address this, we investigated the role of active sensing movements in the refuge tracking behavior ofEigenmannia virescens, a species of weakly electric fish. These fish track the longitudinal movements of a refuge in which they are hiding by swimming back and forth in a single linear dimension. During refuge tracking,Eigenmanniaexhibits stereotyped whole-body oscillations, in addition to its tracking movements, when the quality of the sensory signals degrades. We developed a state-feedback sensorimotor control model of the fish to examine the role of these ancillary movements on the refuge tracking performance of the fish. Here we showed that active sensing movements are tuned to minimize the predictive uncertainty in state estimation, working as a closed-loop controller in parallel with the behavioral task controller. We experimented withN= 3fish to demonstrate the predictive performance of the proposed model, specifically in terms of capturing the characteristics of the smooth-pursuit tracking and active sensing movements of the fish. Our systematic and comparative analysis suggests that the proposed predictive uncertaintybased active sensing generator model produces fish movements that have statistically indistinguishable trajectories from that of the actual fish movements, unlike the recent models present in the literature.